Artificial intelligence in microbial keratitis

The integration of artificial intelligence (AI) into the various aspects of clinical medicine is being explored with heightened interest. Various studies are exploring multiple aspects of microbial keratitis (MK) data involving imaging and diagnostics employing a variety of AI tools. Deep learning (DL) is now seen as a promising diagnostic AI tool for ophthalmic conditions. Its potential as a diagnostic tool in infective keratitis (IK) has been extensively studied in recent times.

The diagnostic accuracy of DL models for keratitis using corneal imaging such as slit-lamp anterior segment photography, vivo confocal microscopy, anterior segment optical coherence tomography, and corneal topography/tomography as diagnostic datasets has been assessed in various studies.[1] Expert consensus, microbiological results, treatment response, and their combination were engaged as reference standards in most studies. Convolutional neural networks (CNNs) seem to be the commonly used primary DL model in most studies, with good diagnostic accuracy to recognize IK and differentiate it from healthy eyes or other corneal pathologies.

CNNs and other DL models have been observed to achieve equivocal diagnostic accuracy with reference standards for KC and FECD. However, the performance in IK and DED has more potential for perfection as concerns of heterogeneous image quality and objective training criteria seem to require further refinement.[2] Its comparable diagnostic accuracy with ophthalmologists seems to be convincing. However, the challenge lies in its ability to distinguish between various types of IK. Conclusions of the recent meta-analysis[1] point toward overestimation of diagnostic accuracy of DL models with internal validation due to overfitting and recommend external validation for establishing the generalizability of DL models.

CorneAI, another DL model, though trained on diffuser slit-lamp images, seemed to perform effectively utilizing smartphone images.[3] The slit-lamp photography (SLP) generative adversarial network (GAN) model StyleGAN2-ADA, when trained on limited real and supplemented synthetic data, was observed to be able to provide a reliable AI-based MK classification system.[4] Given the ease of obtaining good resolution clinical images with smart phones photography, the CNN-based DL model was seen to perform effectively in diagnosing and distinguishing between the subtypes of MK using smartphone-captured images. This will aid in developing novel rapid diagnostic approaches in management of IK.[5] DL models using transfer learning with ResNet50 architecture to classify culture-confirmed keratitis observed significant diagnostic accuracy in distinguishing FK, AK, and NSK and subtyping FK along with providing treatment strategies.[6] A recent study utilizing AI imaging segmentation algorithm base DL model predictions showed good association of stromal infiltrate area and hypopyon with 3-month visual outcome.[7] Another retrospective study[8] that was done to evaluate three novel Vision Transformer (ViT) frameworks in bacterial and fungal keratitis diagnosis used different images ((broad-beam, slit-beam, and blue-light images) of the anterior segment to improve recognition accuracy. The combination of two or more types of anterior segment images was seen to be enhance the diagnostic accuracy of ViT in diagnosis of bacterial and mycotic corneal infection. DeepIK was noted to perform well in internal, external, and prospective datasets outperforming three other algorithms (DenseNet121, InceptionResNetV2, and Swin-Transformer).[9]

In an attempt to evaluate the ability of an automated classification system to surpass the current gold standard in MK diagnosis of corneal smear cultures, the effective automatic classification methods using DL for multitype infectious keratitis diagnosis was studied.[10] The large size of the images of whole slides of potassium hydroxide (KOH) smears precludes the employment of conventional computer vision methods such as CNNs for analysis. An earlier retrospective observational study, hence, evaluated the performance of a DL framework, dual stream multiple instance learning (DSMIL), in automation of whole slide imaging (WSI) of potassium hydroxide (KOH) smears to provide precise and rapid detection in mycotic keratitis.[10] DSMIL uses image segmentation of the WSI to extract relevant features by producing heat maps to visualize areas contributing to the diagnostics prediction and compiles them to enable a comprehensive diagnostics. The competence of DSMIL framework in handling large, high-resolution WSI data along with precise detect fungal infections using heat maps-generated visual explanations seems to be a promising in the automation of KOH smears image interpretations.

The EfficientNet_B0 model succeeded as the best performer among five models (EfficientNet_B0, EfficientNet_V2_S, ResNet50, Vision Transformer (ViT), and DeepIK) that were evaluated MK classification. EfficientNet_B0 model effectively distinguished the normal eyes and four IK types, establishing the role of this DL model in diagnosing MK.[11]

The prospective data set combining both imaging and clinical data of bacterial and fungal keratitis was studied to develop and evaluate multimodal machine learning models for differentiating bacterial and fungal keratitis comparing three prediction models (clinical data model, a computer vision model using EfficientNet architecture, and a multimodal model) in a study from South India recently.[12] The computer vision model outperformed as the best IK classifier suggesting the reliable diagnostic accuracy of image-based DL with the use of prospective, sequentially collected, representative datasets.

Another recent study assessed the implementation of large language models (LLMs) for large-scale MK data analysis, studied extraction of MK descriptors from the clinical notes of hospital electronic medical records (EMRs) adopting the zero-shot prompting to approach, and compared them with the annotations of clinical cornea experts.[13] GPT-4o and GPT-4o mini were prompted to extract the three MK descriptors (centrality, infiltrate depth, and thinning), annotated by human experts, and the agreement measures between the LLM responses and human annotations were analyzed. While good agreement was noted between GPT-4o and GPT-4o mini and human annotations in extracting MK descriptors, the quality and consistency of EMR documentation seem to influence the efficiency and performance of LLM architectures in the detection of MK descriptors in big data analysis.

Comparison of use of real cases, AI-generated images, and real medical images to evaluate diagnostic accuracy of medical students for bacterial, fungal, and herpetic keratitis noted equivocal efficacy and is being advocated as a method of clinical ophthalmology training,[14] enabling the use of AI in medical education methods. While AI modes utilizing image interpretation techniques seem to perform well in pathogen detection in IK, significant limitations such as single retrospective methodology, lack of diverse datasets, restricted generalizability, inadequate classification methods, and lack of uniformity in ground truth detection standards preclude its adoption in clinical practice.[15] The need of the hour with the current scenario seeking to establish the performance efficacy of AI in the various aspects of MK detection and treatment and future research needs to incorporate optimal enhancements with larger datasets to ensure better precision in these aspects.

Better clarity on patient characteristics reporting, engaging datasets of homogeneous study populations, will be more beneficial in future studies on DL models for MK. AI-related black box issues remain a concern, resulting in lack of transparency and hence raising doubts on trust and accountability of its performance. Improved reporting of study characteristics, data diversity, external validation, transparency of AI algorithms, and enhanced reliability and generalizability of DL models will need to be the focus of future evaluations in order to make this a more desirable diagnostic approach that can surpass human abilities.